Research And Application On Object Detection Of Remote Sensing Images Based On Deep Convolution Neural Network

Remote sensing satellite has the advantages of high resolution and fast acquisition speed,which makes the remote sensing image data show an exponential growth.Massive data also puts forward higher requirements for remote sensing image processing technology.How to extract remote sensing image information quickly and accurately has become a crucial part of the development of remote sensing technology.Object detection of remote sensing image is a new technology rising with the development of remote sensing.It has the advantages of wide coverage,long distance and high efficiency.It has important military significance and civil value.Remote sensing image object detection and recognition is one of the most basic tasks in the field of satellite remote sensing image processing.Remote sensing image has the characteristics of large field of view,high background complexity,special angle of view,object rotation,small object and so on.It not only provides more regions of interest,but also brings more complex background information,which brings great challenges to object detection.The task of object detection and recognition in optical remote sensing image is to determine whether there are objects in optical remote sensing image,and to locate and classify them.With the rapid development of artificial intelligence and deep learning,affected by the successful application of deep learning in natural scene image object detection,many scholars try to apply deep learning method to remote sensing image object detection,and remote sensing small object detection based on convolution network has become a development trend.This paper designs a series of multi angle rotation remote sensing small object detection algorithms,the main research content and innovation are as follows:(1)A multi-angle rotation remote sensing small object detection algorithm is proposed.Compared with other existing models,it has the characteristics of less model parameters and simple model structure.In this algorithm,we integrate global attention mechanism into FPN to construct GA FPN module,which is used in remote sensing image detection,proposes the MergeNet network structure,merges the features from different layers,and finally introduces the attention mechanism,so that the model has good detection results in remote sensing rotating small object detection in complex environment.Compared with the existing R2CNN,the map of the algorithm is improved by 6.48%.(2)In this paper,DARTS is applied to the automatic search architecture of remote sensing image by searching the structure of remote sensing data in a differential way.The unit structure search of 5 types of remote sensing image data is carried out on a single card,and the distributed parallel unit structure search of 15 types of remote sensing image data is carried out on 8 gpu cards from 4 servers.Then,the NWPU-RESISC45 and AID dataset is classified based on the architecture.Compared with the existing classic network,it has different degrees of improvement(3)We use DARTS to search the basic unit of remote sensing data set,and combine it with neural structure search NAS-FPN,then integrate it into single-stage object detection rotating RetinaNet network,and propose DARTS-FPN network to detect remote sensing objects.Experimental results show that this method can effectively improve the detection accuracy of multi angle remote sensing small object.(4)The proposed remote sensing object detection algorithms are used in military applications.According to the requirements of military intelligence reconnaissance field,a set of US combat aircraft dataset is independently constructed based on Google Earth.MARNet and DARTS-FPN algorithms are applied to the dataset.The experimental results show that the two remote sensing image object detection algorithms have high detection accuracy on the US combat aircraft dataset,and have certain practical application value.Based on the above research work,this paper systematically studies the multi angle rotation remote sensing small object detection algorithm,and proposes a series of algorithms in the aspects of manually designed detection network and detection network based on differential structure search cell structure.